Raised at #6586 is the fact that the number of folds (i.e. the number of partitions of a dataset) is not always able to equal the number of train-test splits. Also, we now have get_n_splits() in all CV splitters, when some of them have the ability to directly specify the number of splits by n_folds or n_iter parameters.

For consistency, all CV splitters that directly specify the number of splits should now use the parameter name n_splits. If so, should we deprecate or just make this part of the move to 0.18's model_selection?